DEEPFUNC · Using deep neural networks to understand functional specialization in the human visual cortex

Over the last few decades, neuroscientists have identified multiple brain regions that perform distinct, often highly specialized functions such as processing faces, understanding language, and even thinking about what other people are thinking. Despite our increased understanding of the computations performed in these regions, the precise causes and origins of functional specialization in the brain are still a mystery and not accessible to direct experimental approaches.Here, we propose to combine cutting-edge computational modelling, large-scale sampling of naturalistic behaviour and human neuroimaging to overcome these limitations. Focusing on visual perception, we will exploit the latest advances in artificial neural networks to probe three critical aspects of functional specialization in the ventral visual pathway: First, by training networks on natural and artificial visual categories and identifying which features result in functional specialization, we will characterize what it is about a visual category that leads to functional specialization. Second, we will leverage large-scale egocentric datasets of infant and adult visual input to test how visual experience and natural input statistics shape functional specialization during development. Third, we will ask why certain neural features become specialized for high-level visual categories in the human visual cortex in the first place. Critically, for each of these aspects, we will close the loop and directly test and validate its predictions in the human brain.Our project will shed light on functional specialization from a new angle by relating functional specialization to the computational constraints of performing tasks in the real world. Using this novel approach, our project tackles some of the most fundamental questions about the functional organization of the human mind and brain the what, how and why of functional specialization.